Normalization and Standardization | Why to Scale the Features? | ML Basics

Key Takeaways

hi and welcome back the topic of this video is data scaling often raw data comes with features that have vastly different ranges for example attendance rates reflecting the proportion of days an employee attends work out of the total available days job satisfaction ratings ranging from 0 to 10 salary which can range from 0 to infinity and let's say debt amounts which can vary from negative Infinity to positive Infinity see how different the scales are data scaling is the process of transforming our data so that different features are on a similar scale if you decide to train a machine learning model that relies on distance calculations like K nearest neighbors or support Vector machines the features having larger scales will dominate the calculations the same applies to neural networks but in a different way varying scales can seriously affect the training process thus even for powerful neural networks data scaling is essential the two most popular scaling methods used a lot in machine learning are minmax scaling and standardization minmax scaling or normalization scales data into the range of negative - 1 to positive 1 or 0 to 1 the resulting range is fixed and all values are proportionally scaled between maximum and minimum values preserving their relative distances useful for cases that require a fixed range for instance RGB images consist of pixels that range from 0 to 255 by scaling it from 0 to one you maintain a consistent scale across all pixel values however it is very sensitive to outliers let's imagine this data as we can see because of one outlier the whole data is concentrated into a small portion of the range making it harder for models to distinguish between them effectively standardization on the other hand converts the data to have zero mean and one standard deviation so unlike minmax scaling It Centers the data and normalizes the spread standardized scale is not bounded so you may have any value but with most values concentrated in -3 to positive3 range while many sources are indicating that outliers do not affect standardization it is not true it is still sensitive to outliers let's see how I have generated a normally distributed data that looks like this and here is the standardized version of it in each iteration I am going to add an outlier and standardize it to see how the distribution changes compared to the original data as we can see the data is becoming skewed and eventually ends up concentrating into a small portion the selection of these two data scaling methods is often not very critical while in some scenarios one method is preferred over the other one we will not go through it in this video but if you don't want to miss such educational videos follow us if you want to learn more about artificial intelligence subscribe to our channel to be aware of the new videos press the like button and let's discuss AI in the comments section